Tense and Aspect Assignment in Narrative Discourse David K. Elson and Kathleen R. McKeown Department of Computer Science Columbia University {delson,kathy}@cs.columbia.edu Abstract tion to create semantic encodings of short stories. To do this, they construct propositions (predicate- argument structures) through a graphical, menu- We describe a method for assigning English based interface, and assign them to intervals on a tense and aspect in a system that realizes sur- timeline. Figure 1 shows a session in which the face text for symbolically encoded narratives. Our user is encoding a fable of Aesop. The top-right testbed is an encoding interface in which proposi- panel shows the original fable, and the left-hand tions that are attached to a timeline must be real- panel shows a graphical timeline with buttons for ized from several temporal viewpoints. This in- constructing new propositions at certain intervals. volves a mapping from a semantic encoding of The left-hand and bottom-right panels contain au- time to a set of tense/aspect permutations. The tomatically generated text of the encoded story, as encoding tool realizes each permutation to give the system understands it, from different points of a readable, precise description of the narrative so view. Users rely on these realizations to check that that users can check whether they have correctly they have assigned the formal connections cor- encoded actions and statives in the formal repre- rectly. The tenses and aspects of these sentences sentation. Our method selects tenses and aspects are a key component of this feedback. We describe for individual event intervals as well as subinter- the general purpose of the system, its data model, vals (with multiple reference points), quoted and and the encoding methodology in a separate paper unquoted speech (which reassign the temporal fo- (Elson and McKeown, 2010). cus), and modal events such as conditionals. The paper is organized as follows: After dis- cussing related work in Section 2, we describe our 1 Introduction method for selecting tense and aspect for single Generation systems that communicate knowledge events in Section 3. Section 4 follows with more about time must select tense and aspect carefully complex cases involving multiple events and shifts in their surface realizations. An incorrect assign- in temporal focus. We then discuss the results. ment can give the erroneous impression that a con- 2 Related Work tinuous action has ended, or that a previous state is the current reality. In this paper, we consider There has been intense interest in the interpre- English tense and aspect in the generation of nar- tation of tense and aspect into a formal under- rative discourse, where statives and actions occur standing of the ordering and duration of events. over connected intervals. This work has been in both linguistics (Dowty, We describe two contributions: first, a general 1979; Nerbonne, 1986; Vlach, 1993) and natu- application of theories of tense, aspect and inter- ral language understanding. Early systems inves- val logic to a generation context in which we map tigated rule-based approaches to parsing the du- temporal relationships to specific tense/aspect se- rations and orderings of events from the tenses lections. Second, we describe an implementation and aspects of their verbs (Hinrichs, 1987; Web- of this approach in an interactive environment with ber, 1987; Song and Cohen, 1988; Passonneau, a basic sentence planner and realizer. The first re- 1988). Allen (1984) and Steedman (1995) focus sult does not depend on the second. on distinguishing between achievements (when an The purpose of the system is to allow users who event culminates in a result, such as John builds are na¨ıve to linguistics and knowledge representa- a house) and processes (such as walking). More Figure 1: Screenshot of our story encoding interface. recent work has centered on markup languages teractive narrative encoding project. Callaway for complex temporal information (Mani, 2004) and Lester’s STORYBOOK (2002) aims to im- and corpus-based (statistical) models for predict- prove fluency and discourse cohesion in realiz- ing temporal relationships on unseen text (Mani et ing formally encoded narratives; Ligozat and Zock al., 2006; Lapata and Lascarides, 2006). (1992) allow users to interactively construct sen- Our annotation interface requires a fast realizer tences in various temporal scenarios through a that can be easily integrated into an interactive, on- graphical interface. line encoding tool. We found that developing a custom realizer as a module to our Java-based sys- 3 Expressing single events tem was preferable to integrating a large, general 3.1 Temporal knowledge purpose system such as KPML/Nigel (Matthiessen and Bateman, 1991) or FUF/SURGE (Elhadad The propositions that we aim to realize take the and Robin, 1996). These realizers, along with Re- form of a predicate, one or more arguments, zero alPro (Lavoie and Rambow, 1997), accept tense as or more attached modifiers (either a negation oper- a parameter, but do not calculate it from a semantic ator or an adverbial, which is itself a proposition), representation of overlapping time intervals such and an assignment in time. Each argument is asso- as ours (though the Nigel grammar can calculate ciated with a semantic role (such as Agent or Ex- tense from speech, event, and reference time or- periencer), and may include nouns (such as char- derings, discussed below). The statistically trained acters) or other propositions. In our implemented FERGUS (Chen et al., 2002) contrasts with our system, the set of predicates available to the an- rule-based approach. notator is adapted from the VerbNet (Kingsbury and Palmer, 2002) and WordNet (Fellbaum, 1998) Dorr and Gaasterland (1995) and Grote (1998) linguistic databanks. These provide both durative focus on generating temporal connectives, such as actions and statives (Dowty, 1979); we will refer before, based on the relative times and durations of to both as events as they occur over intervals. For two events; Gagnon and Lapalme (1996) focus on example, here are an action and a stative: temporal adverbials (e.g., when to insert a known time of day for an event). By comparison, we ex- walk(Mary, store, 2, 6) (1) tend our approach to cover direct/indirect speech hungry(Julia, 1,∞) (2) and the subjunctive/conditional forms, which they do not report implementing. While our work fo- The latter two arguments in (1) refer to time cuses on English, Yang and Bateman (2009) de- states in a totally ordered sequence; Mary starts scribe a recent system for generating Chinese as- walking to the store at state 2 and finishes walking pect expressions based on a time interval represen- at state 6. (2) begins at state 1, but is unbounded tation, using KPML as their surface realizer. (Julia never ceases being hungry). While this pa- Several other projects run tangential to our in- per does not address the use of reference times (such as equating a state to 6:00 or yesterday), this Diagram Relations Perspective R < E1 Before is an area of ongoing work. E 1 E 2 (1) and (2), depending on the situation, can be R realized in several aspects and tenses. We adapt R = E1 Begin E 1 E 2 R < E2 and extend Reichenbach’s (1947) famous system R of symbols for distinguishing between simple and E1 < R During E 1 E 2 progressive aspect. Reichenbach identifies three R < E2 points that define the temporal position of the R R = E2 Finish event: the event time E, the speech time S, and E 1 E 2 R > E1 a reference time R which may or may not be in- R dicated by a temporal adverbial. The total order- R > E2 After E 1 E 2 ing between these times dictates the appropriate R aspect. For example, the simple past John laughed has the relation E < S. R = E because there is Table 1: Perspective assignment for viewing an no separate reference time involved. The past per- event from a reference state. fect John had laughed [by the end of the play] has the relation E < R < S, in that it describe “the formally encoded the story. In the remainder of past of the past”, with the nearer “past” being R this section, we describe our method for assigning (the end of the play). R can be seen as the tempo- tenses and aspects to propositions such as these. ral focus of the sentence. As Reichenbach does not address events with 3.2 Reference state intervals, we redefine E as the transition (E1..E2) In both snapshot and narration modes, we often attached to the proposition (for example, (2,6) need to render the events that occur at some ref- for Mary’s walk). This definition deliberately as- erence state R. We would like to know, for in- sumes that no event ever occurs over a single “in- stance, what is happening now, or what happened stant” of time. The perception of an instantaneous at 6:00 yesterday evening. The tense and aspect event, when it is needed, is instead created by di- depend on the perspective of the reference state lating R into an interval large enough to contain on the event, which can be bounded or unbounded. the entire event, as in Dowty (1979). The two-step process for this scenario is to deter- We also distinguish between two generation mine the correct perspective, then pick the tense modes: realizing the story as a complete discourse and aspect class that best communicates it. (narration mode) and describing the content of a We define the set of possible perspec- single state or interval (snapshot mode). Our sys- tives to follow Allen (1983), who describes tem supports both modes differently.